(ORDO NEWS) — American researchers have created a metamaterial from which they made analog optical computing cells.
If you direct a beam of light at them, the angle of incidence, polarization or wavelength of which are the elements of a mathematical matrix, then the output will be radiation with changed parameters, representing the solution of a given equation.
The operation of such computers is several orders of magnitude faster than the processors used in electronics, even if these are specialized microcircuits, and not universal ones.
At first glance, in addition to universal ones – desktops (desktop PCs), laptops, smartphones and servers – they can also be highly specialized: for example, network equipment, industrial plant controllers.
Nevertheless, with some exceptions, the architecture of all modern computers is fundamentally universal and multifunctional.
This is what allowed digital computers, as cheaper to manufacture due to the mass and manufacturability of solutions, to displace analog ones.
Moreover, even before the middle of the 20th century, the latter were widely used – from relatively massive mechanical or electronic calculators and highly specialized devices like sights to scientific installations for modeling elementary particles and educational simulators of various equipment.
Needless to say, mechanical watches are still in use – this is also an analog computer that counts seconds, minutes and hours from the moment it starts its work. And every self-respecting engineer should still be able to use the slide rule.
As the packing density of transistors on processor chips grows, the physical performance limits of microcircuits are getting closer.
Therefore, there is an active search for architectural and layout solutions that allow increasing the computing power of microelectronics.
One of the ways is the high specialization of the individual blocks that make up the microprocessors.
Modern chips have separate “sections” responsible, for example, only for processing images and digitized signals or accelerating the work of neural network algorithms.
But their possibilities are also not unlimited.
And given that in many microcircuits for consumer electronics, modules designed for universal tasks occupy a smaller chip area than specialized blocks and cache, we can say that this limit is not far off.
There is a solution, and it is not new – hybrid computers, in which some functions are left to the good old analog modules. Indeed, under certain conditions, they can be radically more effective than digital ones.
But there is a limit here too: an analog computer, even if it is part of a larger one, still remains extremely highly specialized, it is almost impossible to reprogram it for other tasks.
Nevertheless, if one finds such a task, which occurs frequently and in wide-profile applied programs, an analog computer within a microcircuit intended strictly for this task should give a large increase in efficiency.
It is in this direction that the American-Dutch team of researchers went: scientists calculated, simulated and created in the laboratory analog cells capable of solving Fredholm integral equations.
They are often found in signal processing, modeling complex and stochastic systems, as well as in image recognition, generation and processing.
The development is a metamaterial structure that refracts, re-reflects and re-radiates light in a special way. A silicon grating is applied to a layer of aluminum oxide (sapphire glass) – in fact, growths of a complex shape.
It is not random and is determined by which equation the cell will solve – this is an analog operator in the equation.
A filler layer of silicon oxide is placed on top of it, which is covered with the thinnest (15 nanometers) gold film. The latter plays the role of a translucent mirror.
When the beam enters the cell through the gold coating, it is refracted and reflected off the silicon grating. Some photons pass through, some return to the mirror and are reflected back.
Thus, a process is performed that is equivalent to the successive integration of an approximate function, and literally at the speed of light.
Radiation can exit the cell only from the side of the glass, and its characteristics can be measured – they will just be the solution to the equation.
During the experiment, the “calculation” of the approximate solution of the Fredholm integral equation to the target accuracy threshold took about 349 femtoseconds.
This is an order of magnitude less than the switching time of the fastest transistor (1.2 picoseconds) and three orders of magnitude less than the time it takes for one cycle of a desktop processor (about 330 picoseconds at a frequency of three gigahertz).
Such analog optical computers are grouped into arrays on a substrate and can be optimized for various equations. The size of each is 400 by 800 nanometers, the thickness is about 640 nanometers, not counting the sapphire substrate.
There is no talk about the commercial introduction of such elements into microprocessors yet, the technology is crude.
But its prospects are huge: even taking into account the need, in addition to the cells themselves, to place an emitter and a receiver on a chip, the potential savings in chip space is impressive.
One analog optical computer occupies an area of less than fifty transistors, if we take into account the most modern TSMC manufacturing process – five nanometers. And it does the work, albeit strictly in one task, but of the whole chip.
Contact us: [email protected]